ci: extract cuda stage actions + runner_config mapping

[hnyls2002]

Stack on top of #25197 (already merged): pulls every CUDA stage job in pr-test.yml into 3 composite actions and routes install-script selection through a single runner_config lookup key.

Composite actions

• .github/actions/setup-cuda-test-stage — checkout, stage-health / maintenance gates, sgl-kernel wheel download, install. Install script is now looked up by runner_config (no longer a caller input).
• .github/actions/run-test-cuda-suite — run_suite.py invocation with suite name, partition flags, optional --timeout-per-file, continue-on-error flag.
• .github/actions/teardown-cuda-test-stage — coredump upload + venv cleanup.

Mapping

scripts/ci/runner_configs.py is the single source of truth for runner_config → install_script. The composite action shells out to it at runtime. Sanity check in CI: every runner_config referenced by a workflow job must exist in the map.

Caller stub now ~20 lines per stage

stage-c-test-deepep-4-gpu-h100:
  needs: [...]
  if: |
    ...
  runs-on: 4-gpu-h100
  strategy: { ... matrix from partitions JSON ... }
  steps:
    - uses: ./.github/actions/setup-cuda-test-stage
      with:
        runner_config: deepep-4-gpu-h100
        sgl_kernel: ${{ needs.check-changes.outputs.sgl_kernel }}
        pr_head_sha: ${{ inputs.pr_head_sha || '' }}
        git_ref: ${{ inputs.git_ref || '' }}
    - name: Warmup ...   # stage-specific, kept inline
    - uses: ./.github/actions/run-test-cuda-suite
      with: { suite_name, partition_id, partition_size, continue_on_error_flag }
    - uses: ./.github/actions/teardown-cuda-test-stage
      with: { artifact_suffix: ${{ matrix.partition }} }

Warmup commands (8-gpu-h200, deepep-4-gpu-h100, deepep-8-gpu-h200) stay inline in the caller since each has different model URLs. Can be folded into the mapping later if it earns its keep.

Scope

CUDA stage jobs only (13 jobs). stage-a-test-cpu keeps its own bespoke setup (uv pip + protoc + rust cache) since the install path is fundamentally different.

AMD / NPU per-commit + nightly suites unchanged — they still use the legacy register_*_ci(suite=...) path and pr-test-amd.yml / pr-test-npu.yml have their own dispatch.

Equivalence verification (all 14 CUDA stage jobs)

Python script renders each stage's effective job spec from both forms (main inline + PR-C caller stub + _pr-test-stage.yml reusable workflow) and compares after GHA expression simplification (literal ==, ternary && || , format(), && identity) and cosmetic normalization. All 14 stages PASS semantic equivalence; 8 KNOWN intentional changes are flagged separately:

• upload-cuda-coredumps artifact-suffix: ${{ matrix.partition }} now standardized to all stages (some stages were missing it on main)
• Run FA4 jit_kernel tests (SM100+) renamed to generic Run extra pytest, gated by new extra_pytest_path input (same script body)

https://gist.github.com/hnyls2002/4cbb2fe077b9a6db710b5e5eeb025725

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[hnyls2002]

/tag-and-rerun-ci